Output guardrails in Prompt Engineering / GenAI - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Why are output guardrails important in AI models?

Output guardrails help control what an AI model produces. Which of these is the main reason to use output guardrails?

A To increase the size of the AI model

B To make the AI run faster during training

C To ensure the AI outputs are safe and do not cause harm

D To reduce the number of layers in the AI model

💡 Hint

Think about why we want to control what the AI says or does.

Output guardrails are used to keep AI outputs safe, ethical, and aligned with user needs. They prevent harmful or unwanted content.

❓ Predict Output

intermediate

What will this output guardrail code print?

Consider this simple Python function that applies a guardrail to filter out negative numbers from AI output predictions. What does it print?

Prompt Engineering / GenAI

def guardrail_filter(predictions):
    return [x if x >= 0 else 0 for x in predictions]

outputs = [-3, 5, -1, 7]
print(guardrail_filter(outputs))

A [-3, 5, -1, 7]

B [0, 5, 0, 7]

C [3, 5, 1, 7]

D [-3, 0, -1, 0]

💡 Hint

Look at how negative numbers are handled in the list comprehension.

The function replaces negative numbers with 0, so -3 and -1 become 0, others stay the same.

❓ Model Choice

advanced

Which model architecture best supports output guardrails for text generation?

You want to build an AI that generates text but must avoid harmful or biased content. Which model architecture helps best to apply output guardrails?

A Transformer with a safety layer that filters outputs before returning

B Simple feedforward neural network without output filtering

C Convolutional neural network designed for image recognition

D Recurrent neural network without any output constraints

💡 Hint

Think about which architecture allows easy integration of output filters.

Transformers with safety layers can check and modify outputs to enforce guardrails effectively.

❓ Metrics

advanced

Which metric helps measure effectiveness of output guardrails?

After applying output guardrails, you want to check if harmful outputs are reduced. Which metric is best to measure this?

A Model training loss value

B Total number of predictions made

C Number of layers in the AI model

D Percentage of outputs flagged as harmful by a content filter

💡 Hint

Focus on measuring harmful content, not model size or training progress.

Measuring the percentage of outputs flagged as harmful directly shows how well guardrails reduce bad content.

🔧 Debug

expert

Why does this output guardrail code fail to block harmful words?

Look at this Python code meant to block harmful words from AI output. Why does it fail to block the word 'badword'?

Prompt Engineering / GenAI

def block_harmful(text):
    harmful_words = ['badword']
    for word in harmful_words:
        if word in text:
            text = text.replace(word, '****')
    return text

output = 'This is a BadWord example.'
print(block_harmful(output))

A The check is case-sensitive, so 'BadWord' is not matched

B The replace method is used incorrectly and does not change the text

C The harmful_words list is empty, so no words are blocked

D The function returns before replacing words

💡 Hint

Check how the code compares words and the case of letters.

The code checks for exact matches, but 'BadWord' with uppercase letters does not match 'badword'.